Found a prompt template that turns any LLM into a surprisingly

Been messing with prompt structures for creative writing tasks and stumbled on a pattern that consistently produces genuinely funny, context-aware insults instead of the usual "you're stupid" garbage. The trick is framing it as a character study rather than a direct attack.

The prompt structure

You are [SPECIFIC PERSONA: e.g., "a burned-out Victorian ghostwriter who reviews modern fanfiction" or "a 14th-century bard who only speaks in backhanded compliments"].

Target: [DESCRIPTION OF TARGET — habits, quirks, specific behaviors, not identity traits]

Constraints:
- Zero profanity, zero identity-based jabs
- Lean on situational irony, specific observational detail, and tonal whiplash
- Maximum 3 sentences
- End on an unexpected image or callback

Go.

Why this works

Standard "roast me" prompts default to generic put-downs because the model has no observational anchor. Giving it a distinct voice with a built-in worldview (the Victorian ghostwriter hates modern tropes; the bard must be polite) forces lateral thinking. The constraints strip away lazy crutches — no swearing means it has to reach for precision.

Real examples from testing

Persona: Exhausted Michelin inspector reviewing a coworker's lunch choices
Target: Dave reheats salmon in the office microwave daily at 11:47 AM
Output: "Dave treats the breakroom like a crime scene and the microwave like his alibi. That salmon has circled the globe more times than Magellan and still arrives colder than the silence in HR when he walks past. I've seen warmer receptions at a Viking funeral."

Persona: Southern auntie who blesses your heart while filing your flaws
Target: My habit of opening 47 tabs and forgetting why
Output: "Honey, your browser tabs multiply like rabbits in a revival tent — 47 open and not a single one remembers why it came to the party. Your RAM is weeping into its sweet tea. Bless your heart, close a window before the ghost of productivity past haunts your battery life."

Getting usable results

Three variables matter most:

1. Persona specificity — "grumpy editor" fails. "Editor who rejected Hemingway for 'excessive paragraph breaks' and still writes rejection letters in fountain pen" works.
2. Target granularity — "my brother" gives nothing. "My brother who texts 'k' to three-paragraph voice notes and pronounces 'gif' with a hard G" gives gold.
3. Constraint tension — The "no profanity + max 3 sentences" combo forces density. Drop either and quality tanks.

Where it breaks

• Abstract targets ("capitalism," "my anxiety") produce poetic but toothless results
• Personas without internal logic (random adjective stacks) sound like mad libs
• Models below 7B parameters struggle with the tonal whiplash requirement

Practical use cases

Beyond entertainment: this pattern transfers to any constrained creative task — product copy with a specific voice, error messages that don't sound robotic, code review comments that sting without being toxic. The persona-constraint sandwich is a reusable architecture.

Currently building a small prompt library around this structure. Happy to share the JSON if anyone wants to stress-test it against their own weird scenarios.